AI Agent GitHub Digest — 2026-08-31

🌏 中文版

Today's Highlights

Today's thread is "MCP is becoming the standard layer through which agents expose capabilities to each other" — agno v3.0.2 lets Agent, Team, and Workflow publish themselves as named MCP tools with one line of config. GitHub's trending charts, meanwhile, surfaced two Agent Skills libraries aimed at completely different audiences: one packages a senior engineer's standard operating procedures for coding agents, the other packages 200+ research databases for scientists. And oh-my-pi, a one-person fork, out-benchmarked the original Pi by tuning harder — a reminder that the tool-calling format underneath coding agents is still far from settled.

Trending Repos

can1357/oh-my-pi ⭐ 28,455

GitHub · TypeScript + Rust · MIT

• What it is: a hardened fork of Mario Zechner's open-source terminal coding agent "Pi," wired up to 60+ model providers, 31 built-in tools, 14 LSP operations, and 28 DAP (debug adapter protocol) operations, with its core rewritten in Rust (roughly 80k lines).
• Why it's worth a look: the original Pi is already a well-known lightweight coding agent, so omp didn't try to reinvent it — instead it obsessed over the one detail most projects skip: tool-call format. The project's published benchmarks show that swapping in omp's edit/read/grep format for the same model pushes Grok Code Fast 1's task success rate from 6.7% to 68.3%, and cuts Grok 4 Fast's output tokens by 61%. Same model, and just swapping the tool interface produces an order-of-magnitude difference — a direct signal for anyone building their own agent harness.
• Tech Stack: TypeScript CLI/TUI + a Rust core (LSP/DAP engine) + the Bun runtime; installable via curl script, Homebrew, Nix, or Bun across all platforms.
• Getting Started: Low — curl -fsSL https://omp.sh/install | sh or brew install can1357/tap/omp gets you running. PRs are currently open to everyone (previously they required a maintainer vouch).

K-Dense-AI/scientific-agent-skills ⭐ 38,894

GitHub · Python · MIT

• What it is: a library of 163 ready-to-use research skills (cancer genomics, drug-target binding, molecular dynamics, time-series forecasting, and more) plus 100+ scientific databases, packaged so any agent that supports the open Agent Skills standard can install them — formerly known as the Claude-only "Claude Scientific Skills."
• Why it's worth a look: the rename from "Claude-only" to "works with any agent" is itself a signal — the skill ecosystem is moving from being locked to a single agent platform toward an open standard any vendor can consume. The project claims adoption by 190,000 scientists, and it pairs with an open-source, locally-runnable "AI co-scientist" (K-Dense BYOK) that demonstrates a full research workflow — a ready-made starting point for anyone wiring agents into a lab or research team instead of inventorying database APIs from scratch.
• Tech Stack: Markdown/YAML skill definitions plus Python tooling scripts; compatible with Cursor, Claude Code, Codex, and Google Antigravity.
• Getting Started: Low — install per the Agent Skills standard; individual skills can be picked separately.

addyosmani/agent-skills ⭐ 90,900

GitHub · JavaScript · MIT

• What it is: a "senior engineer workflow" skill pack curated by Google Chrome engineer Addy Osmani, packaging best practices across six stages — Define → Plan → Build → Verify → Review → Ship (spec-driven development, TDD, a five-axis code review, and more) — into 25 auto-triggering skills and 9 slash commands.
• Why it's worth a look: most agent skill libraries teach "how to use a particular tool"; this one turns "how a senior engineer actually makes decisions" into a process. /build auto, for example, generates a plan you approve once, then runs every task automatically with tests and commits — but still pauses on failures or high-risk steps for a human call, rather than running fully unattended. Combined with the author's standing in the frontend/performance community, the repo hit 90k stars within a week, a sign that "process" skills — not just "knowledge" skills — are gaining real adoption.
• Tech Stack: Markdown skill definitions bound to slash commands, installed in one line into 70+ agents via the open-source skills CLI.
• Getting Started: Low — npx skills add addyosmani/agent-skills installs everything, or use --skill to pick a single one.

THU-MAIC/OpenMAIC ⭐ 23,594

GitHub · TypeScript · MIT

• What it is: a Tsinghua University project that generates immersive multi-agent courses in one click, having just shipped v1.0.0 on August 27 with a new "Pro workbench" that lets you plan a curriculum, generate pages, and revise materials through a chat interface with an agent.
• Why it's worth a look: this is one of the rare open-source projects that lands multi-agent orchestration in a concrete vertical — education content production — with a complete product loop: upload documents, audio, or video as source material; an agent plans the course structure; 20 built-in skills handle slides, quizzes, and interactive elements; and the result exports as an offline classroom package. Where most multi-agent frameworks still stop at demoing "how a few agents talk to each other," OpenMAIC shows what it looks like to wire multi-agent orchestration into a real content-production pipeline.
• Tech Stack: TypeScript plus the @openmaic/* SDK family (DSL/renderer/importer), with swappable models, media, search providers, and storage backends.
• Getting Started: Medium — the one-click generation mode is simple, but the Pro workbench's agent planning and server-side persistence need extra setup for a storage backend (the project ships a one-command Postgres option).

Notable Releases

agno v3.0.2

Release Notes

• Key changes: MCPConfig.tools now accepts Agent, Team, and Workflow instances, remote proxies, and Toolkit objects directly, publishing each as its own named MCP tool (an agent named chief becomes a tool called chief, instead of requiring callers to invoke run_agent(agent_id="chief")); run()'s metadata resolution order changes to component → session → call-site; new integrations land for the Synthorai model provider, WaveSpeed image/video generation, Serply search, and AtomicMail.
• Breaking Changes: MCPConfig/MCPServerConfig now raise on unrecognized keyword arguments instead of silently ignoring them; BaseRemote.acancel_run gains a required auth_token parameter, so third-party BaseRemote subclasses need to add it; AgentOS(mcp=...), MCPConfig, and default_tools are the new spellings for mcp_server=, MCPServerConfig, and enable_builtin_tools — the old names still work as aliases for now but are slated for removal in 3.1.
• What it means for you: if you're already using agno to expose agents as a service, you can now publish an entire Agent or Team as an MCP tool directly instead of hand-rolling a run_agent wrapper. But check for any custom BaseRemote subclass or code relying on the old MCPServerConfig name before upgrading — go through the changelog line by line, since the new validation will raise at boot rather than fail silently.

Today's Takeaway

I used to think the "Agent Skills ecosystem" story was mainly about how many skills a library ships. But putting K-Dense's rename from "Claude-only" to "works with any agent" next to agno turning the agent itself into an MCP tool side by side, they look like two faces of the same trend — whether it's a skill or the agent itself, everything is converging toward "wrap it behind one standard interface any host can call," rather than each project building its own plugin system.

References

• can1357/oh-my-pi
• oh-my-pi README
• K-Dense-AI/scientific-agent-skills
• scientific-agent-skills README
• addyosmani/agent-skills
• agent-skills README
• THU-MAIC/OpenMAIC
• OpenMAIC README
• agno v3.0.2 Release Notes
• GitHub Trending — Daily